Calibration
ScreenSpot Pro → MMMU-Pro
MMMU-Pro is estimated from ScreenSpot Pro with a Michaelis–Menten curve fitted on 12 models measured on both: y = 1.3909·x / (0.59931 + x), R² = 0.41, cross-validated error 10.2 pp. It is used for 4 estimates.
| Estimated model | ScreenSpot Pro | MMMU-Pro | Source |
|---|---|---|---|
| Holo2-235B-A22B | 70.6% | 75.2% | estimated ± 10.2 pp, low confidence |
| Holo2-30B-A3B | 66.1% | 72.9% | estimated ± 10.2 pp, low confidence |
| Holo2-8B | 58.9% | 68.9% | estimated ± 10.2 pp, low confidence |
| Holo2-4B | 57.2% | 67.9% | estimated ± 10.2 pp, low confidence |
Fitted on
| Model | ScreenSpot Pro | MMMU-Pro |
|---|---|---|
| Claude Opus 4.5 | 45.7% | 70.6% |
| Claude Opus 4.6 | 83.1% | 77.3% |
| Gemini 3.1 Pro | 84.4% | 83.9% |
| Gemini 3 Pro | 72.7% | 81.0% |
| GPT-5.4 | 85.4% | 81.2% |
| LFM3-VL-2B | 55.5% | 45.7% |
| Muse Glimmer 30B | 75.4% | 74.0% |
| Muse Spark | 84.1% | 80.4% |
| Qwen3.5 397B | 65.6% | 79.0% |
| Qwen3.6 Plus | 68.2% | 78.8% |
| Qwen3.7 Plus | 79.0% | 79.0% |
| Qwen3.8 Max | 84.5% | 82.3% |